You don't need to belittle someone else's work. It's a series of articles, and author has 2 more articles that aren't related to articles Ciechanowski wrote at all.
I was also surprised by this sentence. It sounds like this is the author's first attempt at running models locally.
Or maybe the author has been running heavily quantized small models all that time — Gemma 4 gguf he's using is Q4 and only 16 GB. In my experience quants like this tend to perform much worse.
It would be nice if DDG made even a token attempt at making their search not shit. I still use it, but mostly out of habit and because I suspect every alternative is also shit.
Well, lossless over Bluetooth is a mess currently.
Only 1 codec is capable of that — aptX Lossless. Then, your transmitting device, phone / laptop / etc., needs to be compatible with it, and that's often not the case. Samsung and Apple don't support it.
I bought USB-C Bluetooth dongle that I use with my iPhone for that exact reason. It looks janky, but I think it's worth it.
Discord isn't under pressure to implement these measures globally.
I fully agree that they shouldn't be blamed for kick-starting age verification, because governments are pushing for this all around the world. But it's simply ignorant to pretend that Discord isn't helping these governments to normalise this process with their actions. They're also signalling that businesses are willing to comply and that they have all means necessary to do this.
I wouldn't call it a response, but rather another thing that normalises ID verification online. Now all these governments can use Discord as a reference that (1) this is possible at scale and (2) companies are willing to do this.
I don't think that's clear. They're certainly burning money, but that's mostly R&D: salaries and training compute. Once you remove those, it's unclear whether the AI companies would be losing money on just inference.
You can't "remove" costs willy nilly to make a company look profitable. Running an independent ISP is an extremely lucrative business if you gloss over the capex requirements of installing and maintaining the infrastructure.
You don't remove them willy-nilly, but evaluating companies based on their operating costs is standard accounting practice. R&D is not part of operating costs.
Capex for inference is in, capex for training is out for that analysis.
It's one of the key metrics of a business. These are growth businesses, everybody assumes they're going to have a lot more customers in the future than they have today.
If they've got X times as many customers next year than they have right now, training costs per customer gets divided by X, but inference costs per customer stay roughly constant.
So for businesses where they expect X to be large, they care about operating costs a heck of a lot more than R&D.
There was some credible analysis that I don't have the link to which estimated 50% gross margins for OpenAI, largely eaten up by operational expenses. So not awful unit economics, but not good either.
Assuming that's even true, the big asterisk is uncertainty around efficiency gains in the future. The intelligence divided by cost ratio is changing very quickly. It is hard to make confident predictions more than 3 months out.
reply